ACCESS datasets for CMIP6: methodology and idealised experiments

Keywords: ACCESS, climate change, climate data, climate model evaluation, climate simulation, CMIP6, coupled climate model, Earth System Model.

The Coupled Model Intercomparison Project (CMIP), overseen by the World Climate Research Programme (WCRP), endeavours to collate the results of a standardised set of experiments from global climate models developed at climate research organisations across the globe. The coordinated experiment design and implementation, along with the consolidated results of many different global climate models, allows for robust analysis and a deep investigation of the physical processes of the climate system and the biases inherent in numerical climate models. The current phase, CMIP6, consists of a base set of idealised experiments (known as the DECK; Eyring et al. 2016) and historical simulations that modelling centres are required to perform in order to participate in CMIP6. In addition, ~322 different experiments have been designed across the 23 Model Intercomparison Projects (MIPs) endorsed by the World Climate Research Programme¹ (WCRP). These MIPs address a range of investigatory avenues – for example, exploring future climate projections under a range of socioeconomic scenarios (ScenarioMIP; O’Neill et al. 2016), and investigating the effects of removing carbon dioxide (CO₂) from the atmosphere (CDRMIP; Keller et al. 2018).

Since the conclusion of CMIP5, two new iterations of the Australian Community Climate and Earth System Simulator (ACCESS), focusing on global climate, have been developed for the submission of simulation data to CMIP6. ACCESS-CM2 (Bi et al. 2020) simulates the physical climate with an updated atmospheric component, whereas ACCESS-ESM1.5 (Ziehn et al. 2020a) has been designed to simulate a fully-interactive carbon cycle; both models also have updated ocean and sea-ice components compared to the CMIP5-era versions of ACCESS. Although the two ACCESS models use broadly similar modelling components, ACCESS-CM2 uses newer versions of the atmospheric, land and sea-ice components, whereas ACCESS-ESM1.5 contains additional land and ocean biogeochemical components to facilitate the carbon cycle. Additionally, an ocean and sea-ice-only version of ACCESS-CM2 has been developed in partnership with the Consortium for Ocean–Sea-Ice Modelling in Australia (COSIMA) at three resolutions – ACCESS-OM2 (1° ocean), ACCESS-OM2-025 (0.25° ocean) and ACCESS-OM2-01 (0.1° ocean, which was not submitted to CMIP6), all referred to collectively as ‘ACCESS-OM2’ (Kiss et al. 2020). See Table 1 for details on the modelling components of each version of ACCESS. ACCESS has participated in the core CMIP/DECK and nine other MIPs, according to the needs of the Australian modelling community and the strengths of each model version.

Table 1.  Summary of modelling components for the four ACCESS global climate models submitted to CMIP6.

One purpose of this paper is to document key details of the ACCESS submissions to CMIP6. Although there are well-defined protocols for each CMIP6 experiment, the differing complexities of participating models and resources available to each modelling group mean that different choices (and sometimes compromises) are made as to how to meet the CMIP6 experimental specifications. Hence, one component of this paper (Section 1) documents these choices for the ACCESS models, including the spin-up that was performed before experiments were commenced and how the various forcing datasets were applied to each model configuration. A summary of the differences in model versions is provided, along with details of relevant model description and evaluation papers. Section 1 also provides a summary of all the experiments performed for CMIP6 with each ACCESS model, and the ACCESS datasets that are currently available on the Earth System Grid Federation (ESGF, which hosts the CMIP6 submissions across a number of data centres; Balaji et al. 2018). A description of the post-processing pipeline used to conform the model output and metadata to CMIP standards is provided, along with a brief analysis of the computational costs associated with the CMIP6 effort of our modelling group using prescribed metrics. The aim of this section is to provide sufficient information to support the reproducibility of experiments and further analysis of the existing ACCESS datasets.

A second aim of this paper (Section 2) is to provide a brief overview of some of the output from the DECK idealised experiments and related experiments from C4MIP (Jones et al. 2016) and CDRMIP (Keller et al. 2018). These experiments are used to characterise key features of a model’s behaviour such as climate sensitivity, climate drift and overall performance against expected climate responses. The analysis focuses on global mean responses and notes how the simulated climate varies across the two ACCESS coupled models. Global carbon fluxes and related biogeochemical variables are presented for ACCESS-ESM1.5 simulations, with a focus on how changing atmospheric CO₂ and changing temperature affect the behaviour of the carbon cycle. This overview is intended as a springboard and to provide context for further studies, and to compare ACCESS simulations with other CMIP6 models in cases where multimodel studies have already been published.

The paper concludes in Section 3 with some comments on the impact of the ACCESS submissions to CMIP6, whereas Section 3 directs the reader to repositories where the model codes and post-processed data are available.

This section documents the steps required for the ACCESS models to participate in CMIP6, including a summary of the model configurations used, the experiments performed, the application of the forcing datasets for use with the ACCESS model and the post-processing required to meet CMIP6 standards.

For its atmospheric component, ACCESS-CM2 uses the UK Met Office (UKMO) Unified Model (ver. 10.6, UM) GA7.1 configuration (Walters et al. 2019, MetUM-HadGEM3-GA7.1) with a vertical resolution of 85 levels and the GLOMAP-mode aerosol scheme (Mann et al. 2010). This is the same atmospheric configuration and resolution as the HadGEM3-GC31-LL and KACE-1-0-G CMIP6 submissions. ACCESS-ESM1.5 is based on an updated version of ACCESS1.3 (which was submitted to CMIP5; see Bi et al. 2013), and uses a UM (ver. 7.3) configuration that is close to GA1 (Bellouin et al. 2011a, HadGAM2) with 38 levels and the CLASSIC aerosol scheme (Bellouin et al. 2011b). Both are run at ‘N96’ horizontal resolution (1.875 longitude by 1.25 latitude) and make use of the Australian-developed land surface scheme CABLE (Community Atmosphere Biosphere Land Exchange; Kowalczyk et al. 2006, 2013), which is integrated into the UM atmosphere. The terrestrial carbon cycle is implemented in ACCESS-ESM1.5 using CASA-CNP (Wang et al. 2010), within CABLE, including accounting for the impact of nitrogen and phosphorus on the carbon cycle.

Additional components of ACCESS include the LANL CICE model for the sea-ice component (Hunke and Lipscomb 2010), and the NOAA/GFDL Modular Ocean Model (ver. 5, MOM5) for the ocean component at nominal 1° resolution (Griffies 2012). CICE and MOM5 are also the primary components of ACCESS-OM2, with 0.25° version of CICE and MOM5 used in ACCESS-OM2-025. ACCESS-ESM1.5 and ACCESS-OM2 (1° version only) also include the ocean biogeochemical modelling component WOMBAT (Whole Ocean Model of Biogeochemistry And Trophic-dynamics, Oke et al. 2013; Law et al. 2017). A tabulated breakdown of the differences between the ACCESS models used for CMIP6 can be found in Table 1.

More extensive documentation is available elsewhere. Bi et al. (2020) describes the ACCESS-CM2 configuration and the spin-up of the model. Climate drift is assessed and the model performance is evaluated including the mean state of the ocean and aspects of climate variability. The simulated climate is compared with that of ACCESS1.3, as used in CMIP5. The climate of this model version has also been evaluated for an atmosphere-only (amip) simulation by Bodman et al. (2020), looking at both global and Australian metrics. Ziehn et al. (2020a) describes the ACCESS-ESM1.5 configuration including its performance compared with ACCESS-ESM1 (Law et al. 2017). The model is evaluated for climate and carbon cycle stability in the pre-industrial control simulation, whereas the later part of the historical simulation is used to evaluate against observations. Kiss et al. (2020) describes the ACCESS-OM2 configuration and provides extensive evaluation of simulations at three different resolutions. The ACCESS-OM2 configurations used for CMIP6 were updated relative to those described in Kiss et al. (2020) including, most notably, improved topography and updated forcing from JRA55-do ver. 1.3 to ver. 1.4. When WOMBAT was active in ACCESS-OM2, the biogeochemical parameters were identical to those described by Ziehn et al. (2020a) and used in ACCESS-ESM1.5.

Prior to beginning an official CMIP6 pre-industrial control experiment (piControl), a model must be initialised and spun-up under CMIP6 pre-industrial forcing conditions, in order to bring the model climate as closely into balance with the forcing as possible. The spin-up period of ACCESS-CM2 was 950 simulation years, during which several changes were made to the model, including the implementation of CABLE over the original land scheme JULES (Joint UK Land Environment Simulator; Best et al. 2011; Clark et al. 2011) and several bug fixes and tuning changes. The last 100 years of the ACCESS-CM2 spin-up were simulated using the final configuration of the model. The physical climate of ACCESS-ESM1.5 was spun-up over a period of 3000 simulation years with only minor changes and bug fixes applied during this time (with only very minor impacts on the climate trajectory), whereas the biogeochemical processes were integrated over the latter 1000 years of this spin-up. The last 600 years of spin-up represent the final ACCESS-ESM1.5 model configuration. OMIP-2 protocols do not require a spin-up period for the standard ocean-only experiments. Detailed descriptions of code changes and parameter tunings applied to the ACCESS models during their relative spin-up periods can be found in their respective model description papers: Bi et al. (2020) (ACCESS-CM2), Ziehn et al. (2020a) (ACCESS-ESM1.5) and Kiss et al. (2020) (ACCESS-OM2 and ACCESS-OM2-025).

The piControl experiment continues directly from the spin-up, with many experiments then branching from the first year of the piControl – designated year 0950 in ACCESS-CM2, and year 0101 in ACCESS-ESM1.5. The piControl underpins most other experiments by providing a baseline with which to account for residual drift in the climate system that is not part of its response to forcings. The piControl experiment is part of the DECK, which also includes an idealised 1% per year CO₂ increase simulation (1pctCO2), an idealised simulation with an abrupt quadrupling of CO₂ concentration (abrupt-4xCO2), an Atmospheric Model Intercomparison Project simulation with prescribed sea surface temperatures (amip) and a historical simulation over the period 1850–2014. Furthermore, for Earth System Models² there is an additional piControl experiment and historical simulation required in which CO₂ emissions are prescribed, and a fully interactive carbon cycle simulated. In these experiments, esm-piControl and esm-hist, atmospheric CO₂ concentrations are calculated according to CO₂ fluxes between the atmosphere, ocean and land; as opposed to prescribing atmospheric CO₂ concentrations as in the piControl and historical.

In addition to the DECK experiments, each MIP consists of Tier 1 (priority) and Tier 2 experiments. Each modelling group has chosen which experiments to complete with Tables 2–4 detailing the list of experiments performed for CMIP6 (as of April 2022) using ACCESS-CM2, ACCESS-ESM1.5 and ACCESS-OM2 respectively. The experiments are presented by MIP, with the start year referring to the first internal model year of the simulation, which is set to real dates where appropriate (e.g. in the historical simulation). For branched experiments that comprise only a single ensemble member, the branch year is usually the same as its start year (e.g. the ACCESS-ESM1.5 experiment 1pctCO2-cdr begins at year 0241, having branched from 1pctCO2 at the end of year 0240), except where the internal year is set to a real date after branching (most commonly in the case of historical-based simulations; e.g. the ACCESS-ESM1.5 experiment hist-bgc branches from the piControl at year 0161, but the internal start date is set to 1850).

Table 2.  ACCESS-CM2 (CSIRO-ARCCSS) experiments.

Table 3.  ACCESS-ESM1.5 (CSIRO) experiments.

Table 4.  ACCESS-OM2 (CSIRO-COSIMA) experiments.

For experiments with multiple ensemble members, there are two different avenues of ensemble methodology: (i) new ensembles branching from a piControl or esm-piControl, and (ii) child ensembles branching from an existing parent ensemble. For case (i) – new ensembles that branch from a piControl (typically historical simulations) – the following methodology applies: each ensemble member of a child experiment is branched from a progressively advanced year of the piControl (or esm-piControl), and the ‘variant label’ of the child member is advanced by one ‘realisation’ (e.g. r1i1p1f1 and r2i1p1f1 are realisations one and two respectively; see the CMIP6 Controlled Vocabulary³ for details). The branch years of new ensembles are specified, where appropriate, in the right-hand column of Tables 2, 3; branch years are 50 years apart for ACCESS-CM2 and 20 years apart for ACCESS-ESM1.5. In cases where avenue (ii) applies, ensemble members of the child experiment will match those of its parent – for example, ensemble member three (r3i1p1f1) of ssp126 was branched from ensemble member three (r3i1p1f1) of the historical. In these cases – primarily future scenario projections that branch from historical simulations – the ‘variant label’ of the child remains the same as the parent from which it was branched.

A second ACCESS-ESM1.5 abrupt-4xCO2 simulation (r2i1p1f1) was run for 1000 years, for submission to the ongoing LongRunMIP⁴ (Rugenstein et al. 2019), which is not an official CMIP activity. Furthermore, a large ensemble (defined as >10 realisations by Deser et al. (2020)) consisting of 40 members has been simulated with ACCESS-ESM1.5 for the historical and tier 1 ScenarioMIP experiments (ssp126, ssp245, ssp370 and ssp585), along with a 30-member ensemble for the four experiments that were simulated from the CovidMIP activity (which sits within DAMIP and comprises a total of six experiments; Gillett et al. 2016; Jones et al. 2021). These large ensembles are available for use by the SMILE⁵ (Single Model Initial-condition Large Ensemble) community, and will be investigated in more detail in papers currently in preparation. Work is also ongoing to expand the ACCESS-CM2 historical and ScenarioMIP ensembles, with the intention to produce a 10-member ensemble.

Three ACCESS-OM2 experiments (see Table 4) were contributed to OMIP through the OMIP-2 protocol (Griffies et al. 2016; Orr et al. 2017; Tsujino et al. 2020) forced by repeated 61-year (1958–2018) cycles of the Japanese atmospheric reanalysis (JRA55-do, ver. 1.4; Tsujino et al. 2018). The 1° ACCESS-OM2, including the ocean biogeochemistry component WOMBAT, ran both OMIP-2 experiments, omip2 (six cycles) and omip2-spunup (six cycles after conducting 33 cycles as spin-up). The 0.25° ACCESS-OM2-025, without the ocean biogeochemistry component, was also used to run the omip2 experiment (six cycles).

Depending on the experiment, ACCESS requires time-varying forcings, including prescribed monthly solar forcing, greenhouse gases (GHGs), volcanic aerosol optical depth, aerosol chemistry emissions and ozone (Eyring et al. 2016). Land-use and land-cover forcing variations have also been implemented for ACCESS-ESM1.5. Most of the input forcing data were supplied either directly from CMIP6 by input4MIPs⁶ (Meinshausen and Vogel 2016; Durack et al. 2018) or prepared by the UKMO for use in their HadGEM3-GC31-LL CMIP6 submission. Other forcings were determined as specified below.

Idealised experiments in CMIP6, such as 1pctCO2 and abrupt-4xCO2, largely follow the piControl, with historical data used in the historical and related experiments (such as DAMIP; Gillett et al. 2016) and the amip simulation. Other MIPs provide or specify additional forcing data (e.g. ScenarioMIP; O’Neill et al. 2016) or protocols (e.g. RFMIP; Pincus et al. 2016).

In general, we adopted the same forcing files for ACCESS-CM2 as were used for HadGEM3-GC3.1 (Sellar et al. 2020), whereas ACCESS-ESM1.5 required further alterations to input data, including interpolation to the HadGAM2 grid, or the use of separately prepared ancillary data as with land-use change forcings. The forcings used for the PMIP (Kageyama et al. 2018) experiments performed with the ACCESS-ESM1.5 (lig127k and midHolocene) follow the PMIP4 protocols (Otto-Bliesner et al. 2017).

For ACCESS-CM2, piControl (and associated experiments) solar forcing is based on a mean total solar insolation (TSI) value of 1361.0 W m⁻², combined with a monthly time-varying spectral file that provides details for the solar cycle, based on the time-averaged mean conditions of the two solar cycles covering the years 1850–1873, and is derived from the recommended solar datasets for CMIP6 (Matthes et al. 2017).

For ACCESS-ESM1.5, the setup was slightly different. The model spin up was initiated prior to release of CMIP6 protocols, therefore both the spin-up and all idealised experiments, including the piControl were performed using the CMIP5 solar constant of 1365.65 W m⁻², as it was in the CMIP5 versions of ACCESS. ACCESS-ESM1.5 does not utilise full spectral variations.

Historical and future scenario solar variability was based on the historical reconstructions of the 1850–2014 period, and projections of the 2015–2299 period, by Matthes et al. (2017). For ACCESS-ESM1.5, the CMIP6 TSI data were adjusted by a constant 4.9 W m⁻² offset in the historical and future simulations to match the CMIP5-style piControl setup (Ziehn et al. 2020a).

All experiments except PMIP use orbital parameters corresponding to the year 2000. For PMIP experiments, Earth’s orbital parameters (eccentricity, longitude of perihelion and obliquity) were adjusted to the estimated values of the Last Interglacial and mid-Holocene periods (Berger and Loutre 1991; Otto-Bliesner et al. 2017). These orbital parameters modulate the magnitude, and seasonal and latitudinal distribution of the incoming solar radiation.

Stratospheric volcanic aerosol optical depths (AODs), originating from explosive volcanic events, are determined from the zonal mean data of Arfeuille et al. (2014) and Thomason et al. (2018), with aerosol optical properties in the solar (shortwave) and terrestrial (longwave) spectrum. As per the CMIP6 protocols, a climatology of the 1850–2014 period was used for all idealised experiments and future scenarios past 2025, with a smooth transition occurring in 2015–2024 between the 2014 values and the climatology.

ACCESS-ESM1.5 and ACCESS-CM2 use different approaches for specifying stratospheric aerosols, with ACCESS-CM2 utilising AODs as a function of latitude, height and wavelength, as processed by the Easy Volcanic Aerosol module (Toohey et al. 2016). For ACCESS-ESM1.5, we derived AODs over four equal latitude bands from the simple 550 nm AOD data provided for use in CMIP6 (Ziehn et al. 2020a). A small offset was applied to account for the difference in the average AOD from CMIP6 and the CMIP5 data used for the ACCESS-ESM1-5 spin-up and piControl.

The GHG concentrations are included in ACCESS models as globally uniform annual mass mixing ratios. ACCESS-CM2 uses CO₂, methane (CH₄), nitrous oxide (N₂O), CFC12-eq and HFC134a-eq, with the latter two representing groups of chlorofluorocarbons (CFCs) and hydrofluorocarbons (HFCs) (Eyring et al. 2016). ACCESS-ESM1.5 uses CO₂, CH₄, N₂O and separate values for CFC11, CFC12, CFC113, HCFC22, HFC125 and HFC134a, which together account for 98% of the change in the historical radiative forcing (Meinshausen et al. 2017). In the piControl, these values follow: CO₂ = 284.317 ppm, CH₄ = 808.25 ppb and N₂O = 273.02 ppb. The GHG concentrations from Meinshausen et al. (2017) were used for the historical and related experiments (such as amip), whereas for all future scenarios the values from Meinshausen et al. (2020) were used. For PMIP experiments, the concentrations of CO₂, CH₄ and N₂O were adjusted to the estimated concentrations of the mid-Holocene and Last Interglacial (lig127k: CO₂ = 275 ppm, CH₄ = 685 ppb and N₂O = 255 ppb; midHolocene: CO₂ = 264.4 ppm, CH₄ = 597 ppb and N₂O = 262 ppb; Otto-Bliesner et al. 2017).

ACCESS-ESM1.5 was also used for CO₂ emissions-driven simulations (such as esm-historical) in interactive mode, with emissions prescribed by Hoesly et al. (2018). Emissions sources were combined and released from the surface, following interpolation to the ACCESS-ESM1.5 grid and scaling to ensure that total annual emissions agree with the CEDS project.⁷ All other GHG concentrations are prescribed for ACCESS-ESM1.5 emissions-driven simulations.

Both ACCESS-CM2 and ACCESS-ESM1.5 use emissions of black carbon, organic carbon, sulfur dioxide, dimethyl sulfide (DMS) and sea salt for determining optical depths from tropospheric aerosol. Sources of carbon emissions include anthropogenic fossil-fuel and biofuel burning (Hoesly et al. 2018), as well as the forest biomass burning emissions of van Marle et al. (2017), as per CMIP6 protocols. Sulfur dioxide emissions originate from both prescribed anthropogenic (Hoesly et al. 2018) and natural (volcanic degassing, calculated internally by climatology) sources which act as a precursor to secondary sulfate aerosol, along with DMS emissions, which are simulated interactively using prescribed seawater DMS. Sea salt emissions are also simulated interactively in both models. Biogenic and natural emissions are not supplied by CMIP6, and are implemented as in the setup of the UKMO CMIP6 models (Sellar et al. 2020), including a fixed monthly climatology of terpene emissions. Mineral dust is not included in the aerosol schemes of either model, and is simulated separately using the six-bin mass-based scheme of Woodward (2011).

In ACCESS-CM2, aerosol emissions are processed, and their evolution and interactions simulated, by the UKCA GLOMAP-mode module (Mann et al. 2010); whereas ACCESS-ESM1.5 uses the CLASSIC scheme of Bellouin et al. (2011b). Emission files were adopted from the setup of the UKMO GC3.1 model simulations, and were used directly in ACCESS-CM2, but ACCESS-ESM1.5 required regridding prior to use.

Although some CMIP6 models simulate ozone, many prescribe ozone distributions. Both ACCESS configurations apply prescribed monthly ozone concentrations with ACCESS-CM2 using 3-D fields and ACCESS-ESM1.5 using zonal means. This input data originated from the ozone data sets of Morgenstern et al. (2017) for stratospheric ozone, and Checa-Garcia et al. (2018) for tropospheric ozone, which were interpolated from pressure levels to hybrid height coordinates by the UKMO (Sellar et al. 2020). For ACCESS-ESM1.5, further interpolation from 85 to 38 model levels was required.

In ACCESS-CM2 simulations (except for piControl), an additional calculation was applied at each simulation year in order to ensure that the prescribed ozone and model tropopause remain consistent as the tropopause height changes with temperature. The ozone redistribution scheme of Hardiman et al. (2019) redistributes excess tropospheric ozone concentrations into the stratosphere, while conserving global ozone mass. ACCESS-ESM1.5 has much lower vertical resolution about the tropopause and this adjustment would have much less effect and is not included.

ACCESS-ESM1.5 uses a simple land-use scheme which allows for annual changes in tile fractions, and reallocates carbon, nitrogen and phosphorus pools accordingly (Zhang et al. 2013; Ziehn et al. 2020a). Forcing data for changes in vegetation fractions were derived from the Land-Use Harmonisation 2 (LUH2) dataset (Hurtt et al. 2017), which were mapped onto the CABLE plant functional types used in the model.

Nitrogen and phosphorus deposition forcings are used in all ACCESS-ESM1.5 simulations. Nitrogen deposition data are provided by CMIP6 (Jones et al. 2016), whereas phosphorus deposition data (not provided in CMIP6) were adopted from simulations of the predecessor of ACCESS-ESM1.5, ACCESS-ESM1, based on datasets used in Wang et al. 2010).

Sea surface temperatures (SSTs) and sea-ice concentrations are prescribed in the atmosphere-only amip simulations of the DECK, the data for which were derived from the input4MIPs datasets of Durack and Taylor (2017), and interpolated to the UM model grid and N96 land-sea fractional mask by the UKMO.

These data were used in all ACCESS-ESM1.5 amip simulations, and in the first three ACCESS-CM2 amip ensemble members (r[1,2,3]i1p1f1), after which time it was discovered that a mismatch between the ACCESS-CM2 land-sea mask and that used to create the ancillary input files resulted in undefined SSTs at some coastal points, for which the model used a default SST of 273.1 K.⁸ A corrected set of SST and sea-ice ancillary files were used for an additional set of three amip simulations with ACCESS-CM2, which were assigned the variant strings r[1,2,3]i1p1f2. We note here that the amip evaluation by Bodman et al. (2020) included only the initial set of ACCESS-CM2 amip ensemble members (r[1,2,3]i1p1f1). ACCESS-ESM1.5 simulations were not affected.

The SST and sea-ice concentrations were also prescribed in all RFMIP experiments. Climatologies were calculated from the first 50 years of the piControl experiment of each model. These were then interpolated to daily values, and adjusted following the AMIP II boundary condition calculation of Taylor et al. (2000) to ensure that the interpolated monthly means match the original climatology. No issues arose from these ancillary files, because they were internally consistent with each of the models and their associated land-sea masks.

The creation of datasets suitable for publication to the ESGF requires significant resources. An outline of the process for ACCESS datasets is provided here. The raw model output from the UM atmospheric component (along with the CABLE land component output) of ACCESS exists in a binary file format (known as ‘Fieldsfiles’ or the ‘PP-format’⁹); and is converted to netCDF4 using the Python package Iris (Met Office 2020) as a first step in the post-processing pipeline (noting that the first several CMIP6 production simulations used cdms2¹⁰ instead). All other components of ACCESS natively produce model output in netCDF format.

A Python-based software package, referred to as the ‘ACCESS Post-Processor ver. 4.0’ (APP4; Mackallah et al. 2022), has been developed to enable the generation of CMIP6 data using ACCESS model output, and is based on an earlier version of the APP that was used for the ACCESS submission to CMIP5 (Collier and Uhe 2012; Uhe et al. 2012). APP4 uses a python2.7 environment and cdms2 to extract and prepare the relevant fields from the model output, and relies on PCMDI’s ‘Climate Model Output Rewriter’ (CMOR ver. 3.4¹¹; Nadeau et al. 2016) to create final data products, ensuring that they adhere to CMIP6 quality standards and follow the CMIP6 data reference syntax¹² (DRS) and CF metadata conventions.¹³ Note that the model_id for ACCESS-ESM1.5 is ACCESS-ESM1-5, this is reflected in all directory and file names. Primary inputs include an experiment-specific data request (created using the dreqPy tool ver. 01.00.31, Juckes et al. 2020), the CMIP6 CMOR Tables,¹⁴ which define the controlled vocabulary and conventions, and an internally created mapping file that specifies how each CMIP6 variable is generated from the existing fields in the ACCESS model output. Reproducibility and code availability are detailed in Section 3.

Key variables that require significant treatment beyond simple arithmetical calculations include overturning circulations and streamfunctions (e.g. msftyz), ocean and sea-ice transports (e.g. mfo), and sea-ice extents, volumes and areas (e.g. siextentn). These are performed by APP4 internally with Python functions, prior to the use of CMOR. Furthermore, many variables require minor manipulations such as averaging over spatial, temporal or tiled dimensions (e.g. most annual variables are based on monthly data), unit conversions (performed by CMOR where possible), masking or the filling of missing values, and the calculation of climatologies. Basic quality checking of metadata is performed using CMOR’s PrePARE tool, along with automatically generated plots to identify spurious and missing data.

As of 11 April 2022, 37 028 ACCESS-CM2 datasets, 265 166 ACCESS-ESM1.5¹⁵ datasets and 822 ACCESS-OM2 datasets have been submitted to the ESGF,¹⁶ where each dataset is a single version of an individual variable at a given time frequency for each ensemble member. Version strings (e.g. ver. 20191225) contain the date on which the dataset was generated, with new versions being generated in the case of retraction and republication (see the CMIP Errata service for details on all ACCESS dataset retractions¹⁷). There have been DOIs minted for datasets grouped by MIP and the relevant citations are noted in Tables 2–4.

All simulations with ACCESS models for CMIP6 were performed on the High Performance Computing (HPC) systems of the National Computational Infrastructure (NCI) at the Australian National University. The supercomputer Raijin was used for the first several ACCESS simulations performed, including the piControl-spinup and most of the DECK experiments; and for the historical and ScenarioMIP experiments, the first realisation of ACCESS-CM2 and first three realisations of ACCESS-ESM1.5. All later simulations were run on its replacement system Gadi. Tests with ensembles of fifty 1-year simulations showed that the climates simulated on the two machines were indistinguishable. Although not bit-wise reproducible, the results were consistent across all variables to the same level of internal climate variability seen among ensemble members, indicating that the model simulations were consistent across HPC systems.

ACCESS-CM2 simulations were performed in suites developed using the Rose engine¹⁸ and scheduled using the Cylc workflow engine.¹⁹ Most ACCESS-ESM1.5 simulations were configured and scheduled with scripts held at NCI, with the exception of PMIP experiments which were run using the payu framework developed by the ARC Centre of Excellence for Climate Extremes (CLEX). The payu framework was also used for the ACCESS-OM2 simulations. Code availability is detailed in Section 5.

Following an assessment of the computational performance of CMIP6 Earth System Models by Balaji et al. (2017), we present the performance of the ACCESS models against the metrics described by the Computational Performance MIP (CPMIP) project in Table 5. Values are taken from simulations performed on NCI’s Gadi HPC system (Cascade Lake nodes). The CPMIP resolution is the total number of 3-D grid points, and in the coupled model configurations is dominated by the ocean, so ACCESS-CM2 and ACCESS-ESM1.5 have quite similar values. However, the extra complexity of the atmospheric physics means that the atmosphere dominates the computational cost and limits the scaling in ACCESS-CM2. Therefore there is no close relation between resolution and cost in these metrics. Complexity (number of prognostic variables) is larger for ACCESS-CM2 than ACCESS-ESM1.5 because of the more sophisticated atmospheric chemistry and aerosol scheme. It is larger for ACCESS-OM2 compared to ACCESS-OM2-025 because only the former includes the ocean biogeochemical component WOMBAT. For ACCESS-CM2 and ACCESS-ESM1.5, data intensity (measured in gigabytes per compute hour) and data output cost depend on whether sub-daily data are saved from the atmospheric component; values presented are from a simulation in which sub-daily data were not saved.

Table 5.  Computational performance of ACCESS models following CPMIP (Balaji et al. 2017) metrics.

In this section we present a range of metrics from different components of the climate system, both physical and biogeochemical, from the two fully coupled models ACCESS-CM2 and ACCESS-ESM1.5. The ocean-only ACCESS-OM2 results will be presented in a separate paper, in preparation. These metrics show the stability of the models in control experiments and also the responses to perturbations over a span of timescales in idealised climate warming scenarios. The figures that follow demonstrate that the impacts of these perturbations vary widely across the climate system; some components can be sensitive to a particular perturbation while insensitive to others. Some impacts are persistent but others are relatively transient. There is insufficient space here for a detailed study of the simulated fields or their responses to climate change. Rather, the overview here provides context for and motivates further analysis. The mean climate states and drifts in the models are already reported in their respective model description papers (Bi et al. 2020; Ziehn et al. 2020a), so the focus here is on comparing their behaviour. Climate responses from historical and future projections from ACCESS models are not covered here but where there is relevant evaluation as part of a multimodel study, this is cited in the discussions below.

This overview uses only a small selection of the variables submitted to CMIP6 (e.g. Dix et al. 2019a; Ziehn et al. 2019c) and their descriptions are listed in Table 6. The global mean physical climate variables used for ACCESS-CM2 and ACCESS-ESM1.5 are near-surface air temperature (tas), top of atmosphere (TOA) net energy balance down (rsdt − rsut − rlut), sea-ice area (siarean + siareas) and ocean temperature (thetaoga). For ACCESS-ESM1.5, the global biogeochemical (BGC) metrics used are net land carbon flux (nbp, positive into land), land net primary productivity (npp), net ocean carbon flux (fgco2, positive into ocean), ocean productivity (intpp), ocean O₂ flux (fgo2, positive into ocean), ocean surface acidity (derived from dissic, talk and equilibrium constants) and global atmospheric CO₂ mixing ratio (co23D).

Table 6.  List of CMIP6 variables used in the present analysis.

We first discuss the control simulations and then the climate response for the experiments where CO₂ is quadrupled (abrupt-4xCO2) or increases by 1% per year (1pctCO2). Effective radiative forcings calculated using RFMIP experiments (Pincus et al. 2016) are also presented. This is followed by C4MIP (Jones et al. 2016) and CDRMIP (Keller et al. 2018) experiments that are based on the idealised experiments and have been run with ACCESS-ESM1.5. Selected BGC fields are then discussed for the idealised experiments and their variants.

The piControl simulation has been run for 500 years with ACCESS-CM2, and 1000 years with ACCESS-ESM1.5. The longer piControl run for ACCESS-ESM1.5 was required for the long reversibility experiment in CDRMIP and the extended length (1000 years) run for abrupt-4xCO2 (r2i1p1f1). Results from the first 300 years of the piControl of both models are shown in Fig. 1.

Fig. 1.  Global physical climate metrics from CMIP idealised experiments. Trends of anomalies in average global surface temperature (top left), net radiative energy flux (top right), sea-ice cover (bottom left) and anomalies in average total ocean temperature (bottom right), from idealised CMIP6 experiments with ACCESS-ESM1.5 (solid lines for transient experiments, black for piControl and green for esm-piControl) and ACCESS-CM2 (dashed lines for transient experiments, red for piControl). 12-month boxcar filters have been applied to all metrics.

Temperatures are plotted as anomalies relative to the first decade of the respective controls. There is a difference (not shown) of ~0.6 K in the global near-surface air temperature (tas) of the two models at the start of their respective control simulations (years 0950 and 0101 for ACCESS-CM2 and ACCESS-ESM1.5 respectively), with ACCESS-ESM1.5 being warmer; this is likely due to the longer spin-up of ACCESS-ESM1.5. There is a small drift in global tas in both models; a linear fit gives 4.99 × 10⁻⁴ K year⁻¹ for ACCESS-CM2 (500 years) and 6.54 × 10⁻⁵ K year⁻¹ for ACCESS-ESM1.5 (1000 years). These drifts are close to those of global ocean temperatures (thetaoga, Fig. 1d) of 5.51 × 10⁻⁴ K year⁻¹ for ACCESS-CM2 and 6.39 × 10⁻⁵ K year⁻¹ for ACCESS-ESM1.5. The drift in total ocean temperature of ACCESS-CM2 is approximately twice the magnitude of an earlier version of ACCESS, ACCESS-1.3, and has the opposite sign (Bi et al. 2020); the drift is substantially less in ACCESS-ESM1.5 due to the long spin-up. Drifts in coupled models are common, and the ACCESS models’ drifts are similar to other CMIP6 models (Irving et al. 2021).

As reported in Ziehn et al. (2020a), the drift in ACCESS-ESM1.5 does not correspond with the TOA net radiation imbalance (Fig. 1b), which shows an energy loss of 0.25 W m⁻². Approximately 40% of the global energy imbalance is contributed by CABLE, but the source of the remainder is not known (Ziehn et al. 2020a). Bi et al. (2020) discuss the ACCESS-CM2 TOA radiation imbalance in relation to the total ocean warming, noting that 87% of the ocean’s net gain in energy is due to the TOA imbalance, with the remainder attributed to a lack of energy conservation. These discrepancies in energy conservation are similar to other CMIP6 models (Irving et al. 2021). Fig. 1c also shows global sea-ice area which is relatively stable at ~20 × 10¹² m² for the piControl run from both models, with drifts of less than −0.003 × 10¹² m² year⁻¹.

A second control run, esm-piControl, is required for models with an interactive carbon cycle such as ACCESS-ESM1.5. In this simulation the atmospheric CO₂ is not fixed but evolves depending on the net land and ocean carbon fluxes. For a control run, these fluxes should be close to zero (globally and averaged over decades) so that the atmospheric CO₂ does not drift. The 500-year esm-piControl run starts after 170 years of spin-up with freely evolving atmospheric CO₂, with the spin-up starting from the initial year of the standard piControl. The global mean surface-layer CO₂ (Fig. 2) drifts slightly in the esm-piControl run (1 ppm century⁻¹) due to global mean land and ocean carbon fluxes to the atmosphere of −0.03 and 0.05 Pg C year⁻¹ respectively.

Fig. 2.  Global average surface-layer CO₂ from emission driven experiments with ACCESS-ESM1.5: esm-piControl, ZECMIP and CDRMIP pulse experiments. For reference, piControl and 1pctCO2 are included. 12-month boxcar filters have been applied to all metrics.

This drift is well within the guidelines for C4MIP of less than 5 ppm century⁻¹ (Jones et al. 2016). The drift in CO₂ leads to a larger temperature drift (2.26 × 10⁻⁴ K year⁻¹) in the esm-piControl compared to the piControl but this is still very small (as seen in Fig. 1a).

The abrupt-4xCO2 and 1pctCO2 simulations see large increases in atmospheric CO₂, producing an increase in temperature, a reduction in sea-ice area and a positive TOA energy balance. Results from these simulations with ACCESS are shown in Fig. 1, with the greater warming in ACCESS-CM2 over ACCESS-ESM1.5 due largely to the different atmospheric configurations. These simulations can be used to calculate a model’s equilibrium climate sensitivity (ECS, based on abrupt-4xCO2; Gregory et al. 2004) and transient climate response (TCR, based on 1pctCO2). For ACCESS-CM2, the ECS and TCR are 4.7 and 2.1°C respectively (Meehl et al. 2020), whereas ACCESS-ESM1.5 has an ECS of 3.9 and TCR of 2.0°C (Ziehn et al. 2020a), which is similar to the climate sensitivities of ACCESS1.0 and ACCESS1.3 (3.6 and 3.9°C respectively; Zelinka et al. 2020) which were submitted to CMIP5.

Both the mean and spread of ECS values among models submitted to CMIP6 are significantly higher than those in CMIP5 (Grose et al. 2020). Furthermore, several CMIP6 models simulate climate responses that do not match ECS estimates from independent assessments using multiple lines of evidence; for example, Sherwood et al. (2020) determined a likely ECS range of 2.3–4.5 K, based on historical and palaeoclimatological records. This overall increase in CMIP6 ECS is largely driven by a subset of models with new or updated physics, such as ACCESS-CM2, which tend to have high ECS values. An investigation of 50 CMIP6 models by Zelinka et al. (2020), in which 16 were found to have an ECS higher than ACCESS-CM2 (including the Met Office models), demonstrated that these ECS increases are primarily due to the representation of cloud microphysics. This provides an explanation as to why ACCESS-CM2 saw the increase but ACCESS-ESM1.5 did not, since the atmospheric components in ACCESS-CM2 have diverged the most since the ACCESS1.3 configuration in CMIP5. Interestingly, the KACE-1-0-G model (Lee et al. 2020) simulated an ECS of 4.5°C (Meehl et al. 2020), very similar to ACCESS-CM2; these two models share the same atmosphere (UM ver. 10.6) and similar ocean model (MOM4.1 in KACE-1-0-G, and MOM5 in ACCESS-CM2). By contrast, HadGEM3-GC31-LL uses the same atmosphere but a different ocean (NEMO ver. 3.6; Storkey et al. 2018), and simulated a significantly higher ECS (ECS = 5.5°C; Senior et al. 2020). This suggests that the ocean and sea-ice components may play a significant role in ECS; however, further investigation is required to understand this complexity.

The larger climate sensitivity of ACCESS-CM2 also shows up in the larger TOA energy imbalance from the increase in atmospheric CO₂; this is likely due to a GA7.0 update in the representation of CO₂ absorption, resulting in a higher CO₂ effective radiative forcing (ERF; discussed further below in relation to RFMIP). As expected, the larger temperature increase in ACCESS-CM2 leads to a larger loss of sea ice, with rapid initial sea-ice loss in both models for the abrupt-4xCO2 case. ACCESS-CM2 shows a further period of rapid loss around years 40–60, driven by the northern polar region and possibly connected to the higher climate sensitivity.

The two idealised ACCESS-CM2 runs (abrupt-4xCO2 and 1pctCO2) reach very similar levels of minimum sea ice, compared to the two idealised ACCESS-ESM1.5 runs where the 1pctCO2 sea-ice area does not drop as low as in the abrupt-4xCO2. This is likely due to the late-simulation increase in surface temperature in the ACCESS-CM2 1pctCO2 run, which outstrips the ACCESS-ESM1.5 1pctCO2 temperature increase after 80 years (see Fig. 1a), likely a result of its higher ECS. Total ocean temperature (Fig. 1d) increases more slowly than the surface air temperature, particularly in the abrupt-4xCO2 simulation, reflecting the time taken for heat to be transported into the deep ocean.

As described in Pincus et al. (2016), a series of atmosphere-only experiments have been performed, as part of RFMIP, to characterise the ERF due to human and natural forcings to the climate. Thirty-year simulations were performed using climatologies of sea surface temperature and sea ice taken from the piControl run and forced with one of the following: pre-industrial conditions (piClim-control), all present-day anthropogenic forcing (piClim-anthro), present-day GHG concentrations (piClim-ghg), present-day aerosols (piClim-aer), present-day land-use forcing separately (piClim-lu; ACCESS-ESM1.5 only) or quadrupled CO₂ (piClim-4xCO2; which is subsequently scaled by a factor of 0.2266 to be equivalent to the present-day CO₂ concentration of 1.4× pre-industrial CO₂, and is used to separate out the forcing from CO₂ from other greenhouse gases). The ERF values from these experiments, and the scaled 1.4xCO₂ value, are presented in Table 7. ACCESS-CM2 simulated a significantly higher 4xCO₂ ERF than ACCESS-ESM1.5, and very close to the CMIP6 mean of 7.98 ± 0.38 W m⁻² (Smith et al. 2020). This is mostly due to an improvement in the representation of CO₂ absorption in the band 8–13 μm, which was implemented in GA7.0 and adopted into ACCESS-CM2 (Walters et al. 2019); this is also likely a contributing factor to the high ECS of ACCESS-CM2. ACCESS-ESM1.5 simulated a 4xCO₂ ERF of 7.04 W m⁻², similar to that of the CMIP5 models HadGEM2-ES (6.99 W m⁻², which uses a similar atmosphere; Andrews et al. 2012) and ACCESS1.3 (6.75 W m⁻², the predecessor of ACCESS-ESM1.5). The GHG, aerosol, athropogenic and land-use (ACCESS-ESM1.5 only) ERF values for both ACCESS models are similar to those of HadGEM3-GC31-LL, and sit well within the spread of CMIP6 values (Smith et al. 2020).

Table 7.  Effective radiative forcing from RFMIP experiments (W m⁻²).

ACCESS-ESM1.5 has been used to run experiments for C4MIP (including ZECMIP) and CDRMIP that are variants of the 1ptCO2 simulation, or branch from the 1pctCO2 simulation; many of these utilise the interactive carbon cycle. Global average surface CO₂ concentration for emissions-driven experiments are shown in Fig. 2 and global mean tas for all C4MIP and CDRMIP experiments is shown in Fig. 3. The response of the carbon cycle is discussed in Section 2.3. For climate–carbon feedback analysis, 1pctCO2 simulations are performed where only the radiation scheme (1pctCO2-rad) or only the biogeochemistry (1pctCO2-bgc) is subject to increasing atmospheric CO₂. As expected, tas increases in the 1pctCO2-rad experiment (only slightly less than in 1pctCO2) whereas the 1pctCO2-bgc simulation shows only a very small temperature increase. The small differences in temperature response between 1pctCO2-rad and 1pctCO2 and between 1pctCO2-bgc and piControl are due to feedbacks between the carbon cycle and climate, principally through changes in the evolution of leaf area index and surface albedo. The ACCESS-ESM1.5 temperature response is in the middle of the range of CMIP6 models (Arora et al. 2020).

Fig. 3.  Trends in anomalies of average global surface temperature from extra idealised experiments with ACCESS-ESM1.5. For clarity, emission driven experiments (esm-piControl, esm-pi-CO2pulse and esm-pi-cdr-pulse) start at year 0201 rather than 0001. 12-month boxcar filters have been applied to all metrics.

The C4MIP-ZECMIP experiments (Jones et al. 2019) explore the response of the climate to an abrupt shift to zero CO₂ emissions. Simulations start at various points from 1pctCO2, a prescribed concentration experiment, and continue as emissions-driven experiments (with zero emissions). The temperature response is small in all cases, with a slight cooling for the lowest branch, almost no change in temperature for the middle branch and a small warming for the highest branch. Most CMIP6 models show more cooling than ACCESS-ESM1.5 but the spread across models (as given by the standard deviation) is larger than the mean or median cooling (MacDougall et al. 2020).

ACCESS-ESM1.5 has completed three CDRMIP experiments. The reversibility of the climate system is tested in 1pctCO2-cdr by extending the 1pctCO2 simulation for 140 years with atmospheric CO₂ reducing by 1% per year, and then running for 620 years with pre-industrial CO₂ (noting only 220 years are shown in Fig. 3). With similar results to Ziehn et al. (2020d), in which the previous ACCESS-ESM version was used, global mean tas remains ~1 K above the pre-industrial temperature at the end of the CO₂ decrease and only slowly decreases over the following centuries, remaining 0.3 K above pre-industrial after 600 years.

A further two CDRMIP experiments are perturbations to the esm-piControl, and are run in emissions-driven mode. Atmospheric CO₂ is instantaneously increased or decreased by 47 ppm (equivalent to 100 Pg C) to examine the response to the perturbation over 90 years and whether that response is symmetric. The magnitudes of differences in global mean tas are small (<±0.4 K, see Fig. 3), with the positive temperature difference (due to the positive CO₂ perturbation) being somewhat larger and shorter lived than the negative temperature difference from the negative CO₂ perturbation. However, another positive temperature difference occurs in the last decade of the simulation due to variability and a decrease in temperature in the esm-piControl. This suggests that an ensemble (currently planned) is needed to better constrain the temperature response.

Global mean ACCESS-ESM1.5 biogeochemical metrics from CMIP6 idealised experiments are shown in Fig. 4 and 5. Fig. 4 includes DECK experiments (piControl, abrupt-4xCO2 and 1pctCO2) as well as experiments where the atmospheric CO forcing in the 1pctCO2 simulation is applied separately to the radiation or biogeochemical components of the model. Fig. 5 shows experiments that branch from 1pctCO2 (including experiments associated with ZECMIP and CDRMIP) as well as emission-driven experiments (esm-piControl, pulse experiments).

Fig. 4.  Global biogeochemical metrics from idealised experiments: the piControl, abrupt-4xCO2, 1pctCO2, 1pctCO2-bgc and 1pctCO2-rad. Trends are shown of carbon flux into the land (top left), net land primary productivity (top right), carbon flux into the ocean (middle left), primary ocean productivity (middle right), oxygen flux into the ocean (bottom left) and average sea surface pH (bottom right) from idealised CMIP experiments with ACCESS-ESM1.5. Sixty-month boxcar filters have been applied to remove seasonal cycles and reduce interannual variability in all metrics, except ocean carbon flux and pH which were filtered with a 12-month boxcar to better reveal the differences between simulations for these variables.

Fig. 5.  Global biogeochemical metrics from idealised experiments; the piControl, esm-piControl, 1pctCO2-rev, zero-emission and pulse experiments, with the same layout as Fig. 4. For clarity, emission driven experiments (esm-piControl, pulse and pulse-cdr) start at year 0201 rather than 0001.

Both Fig. 4 and 5 show CO₂ flux and productivity from the land (top rows) and the ocean (middle); ocean O₂ flux and acidification are in the bottom rows. Metrics are filtered by 5- or 1-year box-car filters to remove the seasonal cycle and to reduce large interannual variability, in order to better reveal the differences between simulations.

The net land carbon flux (Fig. 4a) is the difference between photosynthesis and respiration, and is generally small (in the time mean) except where atmospheric CO₂ is changing. The land net primary productivity (NPP, Fig. 4b) shows carbon uptake by plants (which is approximately balanced by soil respiration). The abrupt-4xCO2 simulation shows uptake of carbon to the land initially, and a large increase in NPP because photosynthesis responds more quickly to the increased CO₂ than does respiration. Although the net land carbon flux returns close to zero within ~40 years, NPP (Fig. 4b) remains ~18% higher than the control due to the elevated atmospheric CO₂. This indicates that soil respiration is also elevated relative to the control, given the approximately zero net land flux. Increasing CO₂ in the 1pctCO2 experiment produces increasing NPP through most of the experiment, though the rate of increase declines after ~30 years. This leads to a peak in the net land carbon flux after 20–30 years and then a decline back to around zero mean flux after 100 years. This return to zero land flux is unusual compared to most other CMIP6 Earth System Models (Arora et al. 2020), which tend to produce net land carbon uptake throughout the simulation; the multimodel mean land carbon flux from fig. 2 of Arora et al. (2020) is 3.8 Pg C year⁻¹ at year 0140 (when the 1pctCO2 atmospheric CO₂ reaches four times pre-industrial CO₂). Across these 140 years, the cumulative land carbon uptake for ACCESS-ESM1.5 (215 Pg C) is lower than other models (408–1204 Pg C); models with nitrogen limitation give lower carbon uptake (mean 536 Pg C) compared with those without (mean 754 Pg C) (Arora et al. 2020, fig. 4). ACCESS-ESM1.5 has been run with both nitrogen and phosphorus limitation, and Ziehn et al. (2021b) shows that this contributes to the smaller land carbon uptake.

The separate impacts of warming and increased CO₂ on the land carbon fluxes can be seen in the 1pctCO2-rad and 1pctCO2-bgc experiments. After an initial rise of 10 Pg C year⁻¹ over the first 30 years, NPP increases (if slowly) throughout the 1pctCO2-bgc experiment due to increasing CO₂ but no significant change in temperature (Fig. 3). This is in contrast to the 1pctCO2 run in which NPP declines slightly after approximately 120 years. The impact of warming is confirmed by the 1pctCO2-rad experiment, where NPP decreases substantially when warming occurs but the carbon cycle does not experience increasing CO₂. The response in NPP is also reflected in the response of the net land carbon flux. After the initial increase in land carbon uptake, the decrease in net land carbon uptake in the 1pctCO2-bgc case is slower than in the 1pctCO2 case, remaining ~1 Pg C year⁻¹ at the end of the simulation. The 1pctCO2-rad experiment shows an increasing carbon release from the land as expected. While following the general response of other CMIP6 models in these experiments, ACCESS-ESM1.5 experiences less carbon uptake than other models in the 1pctCO2-bgc experiment, and releases more carbon than almost all models in the 1pctCO2-rad experiment (Arora et al. 2020, fig. A1).

The carbon fluxes for the four experiments that branch from the 1pctCO2 simulation (three ZECMIP and CDRMIP reversibility) are shown in Fig. 5. The ZECMIP experiments show a very small decline in land NPP relative to the 1pctCO2 simulation, but this decline is not seen in the net land flux. Given the relatively small changes in temperature and atmospheric CO₂ simulated in each of these branch experiments (Fig. 2, 3), it is to be expected that a land biosphere approximately in balance (as indicated by the near zero net land flux in 1pctCO2) would remain in balance and continue to give approximately zero net land flux.

Where the ZECMIP experiments show the response of the carbon fluxes to a relatively slow decrease in atmospheric CO₂, the CDRMIP reversibility (1pctCO2-cdr) experiment shows the response to a relatively large decrease. The net land flux initially remains close to zero, but from around year 0220 becomes increasingly negative, indicating a source of carbon to the atmosphere. The maximum source occurs around year 0280, coincident with atmospheric CO₂ returning to pre-industrial levels, after which the flux returns to approximately zero over ~30 years. This behaviour is reflected in the land NPP, which initially decreases slowly, then more rapidly, dropping below the pre-industrial NPP around year 0280, before returning close to, but lower than, pre-industrial NPP. The overshoot and return to pre-industrial land fluxes likely results from the slower return of temperature to pre-industrial compared with atmospheric CO₂. The overshoot to below pre-industrial net land carbon flux and NPP is widespread globally but analysis of an earlier, similar simulation by Ziehn et al. (2020d) showed a variation in response timing at regional scale. Responses varied due to the dominant vegetation type in a region, whether a region was driven more strongly by variations in temperature or precipitation and from the interaction of carbon pools with different turnover times. These regional responses are being further explored in a multimodel study in preparation.

Also shown in Fig. 5 are the BGC responses to the CDRMIP pulse experiments. The abrupt increase or decrease in atmospheric CO₂ (Fig. 2) leads to positive or negative pulses of land CO₂ fluxes. When CO₂ is added to the atmosphere, the NPP increases by ~20%, decreasing to ~5% above pre-pulse levels by the end of the simulation. The land takes up 8 Pg C in the year that the pulse of CO₂ is put into the atmosphere, but the large uptake is short-lived. When CO₂ is removed, the decrease in NPP is 15–20% over 6 years before increasing to ~3% below pre-pulse levels. The loss of carbon from the land is 5–6 Pg C for the first 5 years after the negative pulse of atmospheric CO₂. Of the 100 Pg C added to the atmospheric CO₂, ~40 Pg C is taken up by the land with most of the uptake occurring in the first 40 years. When 100 Pg C is removed from the atmosphere, the land loses ~50 Pg C over 40 years but then slowly takes up carbon, giving a loss of ~40 Pg C by the end of the simulation. The relatively small asymmetry in global land carbon response from the positive and negative CO₂ pulses should be confirmed regionally, and the analysis of these simulations would benefit from an ensemble of runs, which are currently planned.

The latter four panels in Fig. 4 and 5 capture various aspects of the ocean biogeochemistry in the idealised experiments run with ACCESS-ESM1.5. The surface carbon flux into the ocean (Fig. 4c, 5c) is driven by the difference in CO₂ partial pressure across the ocean–atmosphere surface, and is largely controlled by physical processes and the changing atmospheric CO₂. Ocean productivity (Fig. 4d, 5d) is mainly constrained by the supply of nutrients into the upper ocean, where solar radiation can drive biological growth. The O₂ flux across the ocean surface (Fig. 4e, 5e) is also related to ocean circulation, as well as temperature. Trends in surface acidity (Fig. 4f, 5f) are dominated by the atmospheric CO₂, with some influence from the existing alkalinity of the surface ocean. The CO₂ flux and ocean acidification show the least variability, being almost a direct response to atmospheric forcing. Other metrics (e.g. productivity and O₂ flux) are more sensitive to other conditions, such as temperature and ocean dynamics, and are more variable.

The largest carbon fluxes into the ocean are in the abrupt-4xCO2 experiment, which imposes a high constant atmospheric CO₂ boundary condition and results in a flux into the ocean that reduces after the initial shock. A high CO₂ flux also occurs in the 1pctCO2 experiment, where the increasing atmospheric CO₂ drives an increasing flux over the first few decades. A maximum flux in 1pctCO2 is reached when the value of the Revelle factor increases (e.g. Jiang et al. 2019), reducing the efficiency of CO₂ uptake into the ocean. Mean ocean uptake over the last 30 years of the simulation is 5.3 Pg C year⁻¹ which is slightly larger than the CMIP6 multimodel mean of 5.0 Pg C year⁻¹ (Arora et al. 2020, fig. 2).

In experiments that branch from 1pctCO2 (i.e. ZECMIP and 1pctCO2-cdr), the ocean takes up CO₂ relatively strongly at the branch points (Fig. 5c). In the ZECMIP experiments this uptake decreases after the change to zero emissions, initially rapidly and then more slowly. This is driven by the stabilisation of the atmospheric CO₂ and the reduction in the difference between ocean and atmospheric CO₂ partial pressures. In 1pctCO2-cdr, the ocean changes from a sink to a source after 30 years because the atmospheric CO₂ drops below the average partial pressure of CO₂ in the surface ocean. The source increases in magnitude until atmospheric CO₂ stabilises at pre-industrial levels, after which the ocean flux slowly returns towards zero.

Also shown in Fig. 5 are the BGC responses to the CDRMIP pulse experiments. The abrupt increase or decrease in atmospheric CO₂ (Fig. 2) leads to positive or negative pulses of ocean carbon fluxes, similar to those seen in the land carbon fluxes. The response is reasonably symmetric for the ocean flux, with the pulse decaying over ~20 years. Of the 100 Pg C added to or removed from the atmosphere, the ocean carbon absorption or release is ~30 Pg C by the end of the run, with 50% of the total absorption or release occurring in the first 5 years.

Average ocean productivity is 33.5 Pg C year⁻¹ in the piControl. Initial reductions in productivity of the order of ~2 Pg C year⁻¹ are evident in both the abrupt-4xCO2 and 1pctCO2 experiments (Fig. 4d). Interestingly, these two experiments show an increase in productivity after their initial decrease and, by the end of the abrupt-4xCO2 case, global productivity becomes greater than that of the piControl experiment. There is considerable uncertainty in the productivity response from high-end climate change scenarios; the standard deviation between CMIP6 models of changes in global ocean productivity in ssp585 experiments is greater than the multimodel mean (Kwiatkowski et al. 2020). Preliminary analysis shows that ocean productivity changes are not a simple, uniform global response, but rather that local productivity can increase or decrease in different regions driven by regional changes in temperature, mixed layer depth or nutrient supply (not shown in the global analysis).

In 1pctCO2-bgc, the CO₂ fertilisation experiment, there is negligible impact on productivity and O₂ fluxes compared with 1pctCO2, whereas in 1pctCO2-rad productivity and O₂ flux respond in the same way as 1pctCO2. There is no CO₂ fertilisation process in the ocean biogeochemistry component of the ACCESS-ESM1.5. Hence, productivity in 1pctCO2-bgc follows the piControl, and in 1pctCO2-rad follows 1pctCO2.

The O₂ flux trends (Fig. 4e, 5e) respond to changes in ocean circulation, temperature and productivity but, unlike CO₂ fluxes, there is no change in the direct atmospheric forcing for O₂. There is a non-zero O₂ flux in piControl (~−76.4 Tmol O₂ year⁻¹) due to the remineralisation of detritus in regions with zero O₂ in the model, as discussed in Ziehn et al. (2020a). In abrupt-4xCO2, the balance in the O₂ flux is broken by rapid stratification of the ocean, increasing the vertical stability in the ocean and shutting down the uptake of O₂. There is an increase in O₂ outgassing in 1pctCO2 and 1pctCO2-rad as both stratification and surface temperature rise, which is equivalent to the deoxygenation of the ocean in climate change scenarios of ACCESS and other CMIP6 models discussed by Kwiatkowski et al. (2020). The O₂ outgassing reduces in all experiments branching from 1pctCO2 (Fig. 5e). In 1pctCO2-cdr, O₂ flux reaches parity with piControl after ~100 years, followed by a relative influx for ~50 years until atmospheric CO₂ is reduced to pre-industrial levels. This late influx of O₂ is small relative to the outgassing, such that the total ocean O₂ content is reduced by ~5% relative to the pre-industrial content.

Ocean acidity is not a direct model output of ACCESS-ESM1.5, but can be diagnosed from other ocean carbon variables; the metric previously has been used in model intercomparisons of stresses on marine ecosystems (e.g. Bopp et al. 2013; Kwiatkowski et al. 2020). Acidity is calculated from dissolved inorganic carbon, alkalinity and equilibrium solubility constants. These constants are functions of temperature and salinity, and are determined with the same subroutines used to calculate air–sea CO₂ fluxes within the ocean model. Trends in surface ocean acidity (Fig. 4f, 5f) closely follow the atmospheric CO₂ boundary condition (Fig. 2), with minimal variability. Notable from the trends shown, and particularly for the branching experiments (ZECMIP and CDRMIP-pulse experiments), is that ocean acidification and atmospheric CO₂ are slow to recover from their perturbed states, certainly beyond the 100-year span of the experiments shown. Trends in surface alkalinity due to changes in ocean structure and circulation have a small influence on average acidity as seen in the difference between 1pctCO2-rad and piControl, and that acidity in 1pctCO2-cdr does not return to pre-industrial values.

The CMIP6 submission of ACCESS simulation data is much more extensive than the CMIP5 submission, which utilised the models ACCESS1.0 and ACCESS1.3 (Dix et al. 2013). A number of factors contributed to this increase, including the greater scope of CMIP6, the additional capability of the models (ACCESS-ESM1.5 is the first Australian submission with carbon cycle capability), and sufficient compute and data storage resources at the modelling group’s disposal. Compute resources were larger than originally planned due to the awarding of an Australasian Leadership Computing Grant, which enabled DAMIP simulations and an increase in ensemble size for historical and ScenarioMIP experiments. The CMIP6 submission is a major achievement for a relatively small core team especially since the models used for CMIP6 differed more from each other than the two model versions used for CMIP5 and hence had different requirements for their experimental set-up. The OMIP submission would not have been achieved without the COSIMA contribution.

While the ACCESS-ESM1.5 submission was primarily undertaken for the purpose of including Earth system components, the different atmospheric configuration and consequently lower climate sensitivity (compared to ACCESS-CM2) provides good opportunities for comparative studies of the climate generated by each version. The lower compute resource requirement for ACCESS-ESM1.5 has also facilitated additional model use, such as the university-led palaeoclimate simulations contributed to PMIP (Yeung et al. 2021; Choudhury et al. 2022).

It is difficult to fully track the uptake and use of CMIP6 datasets. The Earth System Grid Federation provides statistics of dataset publications and downloads,²⁰ noting that in this context each variable for each experiment is counted as a separate dataset, with further versions also counting as additional datasets. These statistics show that there have been over 51 million downloads of the ACCESS datasets (accessed 11 April 2022) since the first datasets were submitted in December 2019. These download numbers are comparable to models from major international modelling groups, with ACCESS-ESM1.5 in the top 10 models using this download metric.

The number of publications using ACCESS CMIP6 datasets is increasing. The ACCESS-CM2 AMIP simulations are the basis of a paper by Bodman et al. (2020), and an evaluation of historical climate variability and change, as simulated by the ACCESS models, is presented in Rashid et al. (2022). Ziehn et al. (2021b) explores land carbon-concentration and carbon–climate feedbacks using ACCESS-ESM1.5 simulations, whereas Holmes et al. (2022) provides an analysis of ocean warming in ACCESS-CM2. ACCESS data have also been used in many multimodel analyses (e.g. Watterson et al. 2021; Huang et al. 2022). CMIP6 publications are tracked at the CMIP Publication Hub,²¹ which lists 208 publications (as of 8 April 2022), though tracking is dependent on authors adding their publication details, optionally with the models that have been included in their analysis. The number of publications that list the inclusion of ACCESS-CM2 or ACCESS-ESM1.5 is comparable to the number listed for other well-respected models. Papers that have used ACCESS-OM2 code or data are also tracked independently,²² indicating 35 such publications.

ACCESS-CMIP6 data are also now being used as forcing inputs for regional downscaling activities, including CORDEX-CMIP6²³ (Gutowski et al. 2016), COWCLIP²⁴ and Australian national and state climate projections (e.g. Di Virgilio et al. 2022). Further use of the ACCESS models and their CMIP6 datasets is welcomed, with information on the availability of ACCESS software and data provided in the next section.

Owing to licensing restrictions, we cannot provide either the source code or documentation papers for the UM, which is used under a license and collaborative partnership. For information on obtaining a license for the UM go to http://www.metoffice.gov.uk/research/collaboration/um-collaboration. CABLE is distributed under an open source license. It is hosted at NCI, and registration is required to access the code repository. Details can be found at https://trac.nci.org.au/trac/cable/wiki/CABLE_Registration. MOM5 is available to download from https://github.com/mom-ocean/MOM5 (ACCESS-CM2) and https://github.com/COSIMA/ACCESS-ESM1.5-MOM5 (ACCESS-ESM1.5, with WOMBAT). CICE is available at https://github.com/COSIMA/cice5 (ACCESS-CM2, ACCESS-OM2) and https://github.com/COSIMA/cice4 (ACCESS-ESM1.5). The OASIS3-MCT coupler is available at https://github.com/COSIMA/oasis3-mct. ACCESS-OM2 code is available at https://github.com/COSIMA/access-om2 and the version at the time of OMIP submission is available at https://github.com/COSIMA/access-om2/releases/tag/2020-11-12. The ACCESS-OM2 configurations used for CMIP6 submission include the following: https://github.com/COSIMA/1deg_jra55_iaf/tree/omip2 and https://github.com/COSIMA/1deg_jra55_iaf/tree/omip2spunup for ACCESS-OM2; and https://github.com/COSIMA/025deg_jra55_iaf/tree/omip_amoctopo_cycle1 through to https://github.com/COSIMA/025deg_jra55_iaf/tree/omip_amoctopo_cycle6 for ACCESS-OM2-025. Rose/cylc suits, which each contain the experiment configuration of a single ACCESS-CM2 realisation, are held in a revision-controlled repository service hosted at the UKMO. A mapping between suite names and CMIP6 ensemble members can be found at https://confluence.csiro.au/display/ACCESS/CMIP6+Archive+-+ACCESS-CM2. Payu configurations for the ACCESS-ESM1.5 experiments piControl, historical and PMIP can be found at the CLEX GitHub repository https://github.com/coecms (e.g. https://github.com/coecms/esm-lig). Configurations for ACCESS-OM2 can be found at the COSIMA GitHub repository https://github.com/COSIMA. The APP4 post-processing code used to CMORise ACCESS model output is available at https://git.nci.org.au/cm2704/APP4 (Mackallah et al. 2022). The mapping between ACCESS output fields and CMIP6 variables is specified in the file APP4/input_files/master_map[_om2].csv. Both python2 and cdms2 will soon be no longer supported, therefore reproducibility is enabled and ensured through a Conda environment. Future development and access to this tools will be facilitated by the ACCESS-NRI.²⁵ CMORised simulation data are published on the Earth System Grid at Dix et al. (2019a) (ACCESS-CM2), Ziehn et al. (2019c) (ACCESS-ESM1.5), Hayashida et al. (2021) (ACCESS-OM2) and Holmes et al. (2021) (ACCESS-OM2-025). See also https://esgf.nci.org.au/search/cmip6-nci (ESGF data portal, NCI node) and https://doi.org/10.25914/5e6acd0492b39 (NCI GeoNetwork record). NCI users can also access the data directly from project fs38 (https://my.nci.org.au/mancini/project/fs38). Model output, the semi-processed simulation data which include many variables not requested by CMIP6, that has not been CMORised is also available for NCI users in project p73 (https://my.nci.org.au/mancini/project/p73).

CMORised simulation data are published on the Earth System Grid at Dix et al. (2019a) (ACCESS-CM2), Ziehn et al. (2019c) (ACCESS-ESM1.5), Hayashida et al. (2021) (ACCESS-OM2) and Holmes et al. (2021) (ACCESS-OM2-025). See also https://esgf.nci.org.au/search/cmip6-nci (ESGF data portal, NCI node) and https://doi.org/10.25914/5e6acd0492b39 (NCI GeoNetwork record). NCI users can also access the data directly from project fs38 (https://my.nci.org.au/mancini/project/fs38). Model output, the semi-processed simulation data which include many variables not requested by CMIP6, that has not been CMORised is also available for NCI users in project p73 (https://my.nci.org.au/mancini/project/p73).

Josephine Brown is an Associate Editor for the Journal of Southern Hemisphere Earth Systems Science but did not at any stage have editor-level access to this manuscript while in peer review, as is the standard practice when handling manuscripts submitted by an editor to this journal. The authors have no further conflicts of interest to declare.

ACCESS simulations undertaken with the assistance an Australasian Leadership Computing Grant by sponsor, the National Computational Infrastructure (NCI Australia) (grant #12840 under ALCG2020, NCI project gy57). NCI Australia had a governance role in the publication of ACCESS data to CMIP6, and NCI staff contributed to the preparation of this manuscript. COSIMA and A. E. Kiss were funded by Australian Research Council grants LP160100073 and LP200100406. A. E. Kiss was also funded by the Australian Government's Australian Antarctic Science Program grant 4541. This research was undertaken with the assistance of resources from the NCI, which is supported by the Australian Government.